mnist_classifier.py

# Copyright 2018 The JAX Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""A basic MNIST example using JAX with the mini-libraries stax and optimizers.

The mini-library jax.example_libraries.stax is for neural network building, and
the mini-library jax.example_libraries.optimizers is for first-order stochastic
optimization.
"""


import itertools
import time

import datasets
import jax.numpy as jnp
import numpy as np
import numpy.random as npr
from jax import grad, jit, random
from jax.example_libraries import optimizers, stax
from jax.example_libraries.stax import Dense, LogSoftmax, Relu

import jax_scalify as jsa


def loss(params, batch):
    inputs, targets = batch
    preds = predict(params, inputs)
    return -jnp.mean(jnp.sum(preds * targets, axis=1))


def accuracy(params, batch):
    inputs, targets = batch
    target_class = jnp.argmax(targets, axis=1)
    predicted_class = jnp.argmax(predict(params, inputs), axis=1)
    return jnp.mean(predicted_class == target_class)


init_random_params, predict = stax.serial(Dense(1024), Relu, Dense(1024), Relu, Dense(10), LogSoftmax)

if __name__ == "__main__":
    rng = random.PRNGKey(0)

    step_size = 0.001
    num_epochs = 10
    batch_size = 128
    momentum_mass = 0.9

    train_images, train_labels, test_images, test_labels = datasets.mnist()
    num_train = train_images.shape[0]
    num_complete_batches, leftover = divmod(num_train, batch_size)
    num_batches = num_complete_batches + bool(leftover)

    def data_stream():
        rng = npr.RandomState(0)
        while True:
            perm = rng.permutation(num_train)
            for i in range(num_batches):
                batch_idx = perm[i * batch_size : (i + 1) * batch_size]
                yield train_images[batch_idx], train_labels[batch_idx]

    batches = data_stream()

    opt_init, opt_update, get_params = optimizers.momentum(step_size, mass=momentum_mass)

    @jit
    @jsa.scalify
    def update(i, opt_state, batch):
        params = get_params(opt_state)
        return opt_update(i, grad(loss)(params, batch), opt_state)

    _, init_params = init_random_params(rng, (-1, 28 * 28))
    opt_state = opt_init(init_params)
    itercount = itertools.count()
    # Convert weights + optimizer state to scaled arrays (assuming unit scaling initialization).
    opt_state = jsa.as_scaled_array(opt_state, scale=np.float32(1.0))

    print("\nStarting training...")
    for epoch in range(num_epochs):
        start_time = time.time()

        for _ in range(num_batches):
            batch = next(batches)
            # Convert batch to ScaledArray (assuming proper normalized data).
            batch = jsa.as_scaled_array(batch, scale=np.float32(1.0))
            opt_state = update(next(itercount), opt_state, batch)
        epoch_time = time.time() - start_time

        params = get_params(opt_state)
        # Evaluate model without scaling.
        params = jsa.asarray(params)
        train_acc = accuracy(params, (train_images, train_labels))
        test_acc = accuracy(params, (test_images, test_labels))
        print(f"Epoch {epoch} in {epoch_time:0.2f} sec")
        print(f"Training set accuracy {train_acc}")
        print(f"Test set accuracy {test_acc}")